Using DDRNet to enhance VSLAM stability in a moving scene

The paper illustrates the possibility of using a deep dual-resolution neural network (DDRNet) for solving the problem of real-time detection of moving objects (pedestrians, cyclists, cars, etc.) in images. The paper also presents a DDR-SLAM software solution in the field of VSLAM (Visual Simultaneous Localization and Mapping) in dynamic environments. This VSLAM solution is, in fact, an improved implementation of the ORBSLAM2 framework, in which the DDRNet neural network is used to mask moving objects. The results of experimental testing of DDR-SLAM on image sequences from the KITTI and Cityscapes datasets demonstrate that DDR-SLAM ensures low positioning error values in the following scenarios: vehicle movement in dense traffic; long-term stay of a vehicle in a traffic jam; slowing down the movement of a vehicle before an intersection, to which other vehicles are also approaching, etc. It is also shown that DDR-SLAM confidently outperforms the VSLAM solution ORB-SLAM2 in terms of positioning accuracy in scenarios where there is a large number of moving objects. The source code of DDR-SLAM is available at the link (https://github.com/YznMur/DDRSLAM.git).